Generative foundation models for agricultural image annotation: Semi-arid rangeland dataset and reproducibility code

This repository contains the image dataset, reference proxy masks, generative reconstructions, model-derived binary vegetation masks, and analysis code used in the study “Generative foundation models for agricultural image annotation: Reliability assessment in semi-arid rangelands.” The dataset comprises 98 proximal RGB field images representing 10 indigenous forage shrub and rangeland vegetation classes collected in the arid and semi-arid lands of Marsabit County, Kenya. The images represent heterogeneous field conditions, including exposed soil, dry litter, shadows, sparse and dense canopies, woody stems, open crowns, partial occlusion, and visually complex vegetation–background boundaries. Each source image was processed using four deployed generative image systems: gpt-image-2 gpt-image-1.5 gemini-3-pro-image qwen-image-edit-2511 A common prompt constrained reconstruction protocol was used to preserve the target vegetation while reducing visually distracting background information. The archive therefore contains 392 reconstructed image outputs together with corresponding binary vegetation masks. Reference proxy masks associated with the original RGB images are also provided. These masks were generated from the original imagery using a SAM 3.0 annotation workflow and were used as consistent comparative references for evaluating reconstruction-derived vegetation priors. They should be interpreted as reference proxy masks rather than absolute pixel-level ground truth. The accompanying analysis notebook provides a reproducible implementation of the study's multi metric reliability assessment. It includes: Dice and Intersection over Union agreement foreground area difference boundary F1 and boundary IoU structural and skeleton based measurements foreground loss and generated foreground background leakage internal canopy gap inpainting colour and appearance alteration pairwise model agreement multi model consensus analysis model disagreement mapping human review priority ranking The consensus analysis identifies vegetation regions consistently retained across models and regions where the reconstructed masks diverge. These disagreement regions are intended as candidate signals for human review, rather than calibrated predictive uncertainty or validated error maps.Repository structure:1_Original_images_classified2_original_masks_SAM33_masked_images_generated_ChatGPT4_ChatGPT_binary_masks5_masked_images_generated_Copilot6_Copilot_binary_masks7_masked_images_generated_Gemini8_Gemini_binary_masks9_masked_images_generated_Qwen10_Qwen_binary_masksmultimodal_generative_AI_Zenodo.ipynbREADME.mdrequirements.txtCITATION.cff All images and masks used for paired quantitative analysis were standardised to a spatial grid of 1184 × 864 pixels. The included notebook automatically validates the expected folder structure, performs reference v/s model and model v/s model comparisons, generates consensus statistics, and exports reproducible CSV results and figures. This repository is intended to support reproducibility, benchmarking, and further research on generative foundation models, agricultural image annotation, rangeland vegetation analysis, and human-in-the-loop annotation workflows. Authors: Shubham Rana, Oliver Hensel, Abozar NasirahmadiAffiliations: University of Kassel, Germany; Swedish University of Agricultural Sciences, SwedenProject: InfoRange — Increasing Efficiency in Rangeland-Based Livestock Value Chains through Machine Learning and Digital Technologies.AcknowledgmentsThis research was conducted within the InfoRange project, funded by the German Federal Ministry of Research, Technology andSpace (BMFTR) under grant “01LL2201B”. We thank the funding bodies for their support, Compwiz Creations (https://compwiz.io/) for assisting with app development used in data collection, and our German and Kenyan InfoRange partners for their contributions to image collection. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.